Large language models tracked depression severity from brief daily diaries with clinical-level correlation
Closely monitoring depression across days has always been a measurement problem as much as a clinical one. Traditional self-report surveys can miss the texture of symptom change, especially when they are too sparse, too repetitive, or too detached from how patients actually speak about mood, sleep, energy, and function in daily life. In this study, large language models were tested as a way to score depression from brief video diary transcripts, and the resulting ratings tracked both self-report and interview-based symptom severity with notable consistency.
The cohort included 108 participants selected for mental health treatment status, with 45% already in treatment. Each participant completed a clinical interview and then recorded 14 to 21 days of depression surveys and video diaries. Those diary transcripts were then converted into daily depression scores by LLMs, creating a direct comparison between machine-derived ratings, conventional self-report, and clinician-anchored assessment. The design matters because it moves beyond a one-time snapshot and asks whether language models can follow depression as it unfolds from day to day.
The strongest signal appeared within person, where LLM ratings converged with self-report depression at r = 0.45 with a 95% credible interval of 0.40 to 0.49. When ratings were averaged across days to produce an overall depression score, the association strengthened to r = 0.61 with a 95% credible interval of 0.43 to 0.75. Averaged LLM ratings also correlated with depression symptom severity from the clinical interview at r = 0.53 with a 95% credible interval of 0.35 to 0.68. Those are not trivial overlaps. They suggest the model was not merely echoing isolated words or obvious mood labels; it was capturing a broader semantic pattern tied to clinically meaningful distress.
A key finding emerged in how closely the LLM and self-report profiles lined up across daily functioning variables. The profile correlation was 0.98 within person and 0.94 between person, essentially indicating that the two methods moved together in almost the same direction when functioning changed. In a field where measurement mismatch can obscure symptom dynamics, that degree of alignment is striking.
Daily language carried information beyond conventional self-report
The analysis did not stop at correlation. Multivariable regression showed that LLM ratings predicted daily functioning variables with partial betas ranging from 0.12 to 0.28 and predicted interview-rated depression with a partial beta of 0.33 and a 95% credible interval of 0.07 to 0.58, even after accounting for self-reports. In plain clinical terms, the model retained signal after the usual questionnaire data were already in the equation. That suggests the transcripts contained clinically relevant information that structured surveys can miss: nuance in phrasing, hesitation, self-evaluation, and the lived context around symptoms.
This is where the biological and behavioral logic becomes important. Depression is not only a count of endorsed symptoms; it is also a disorder of how people describe their experience, weigh their energy, and frame their ability to function. Free speech can capture subtle shifts in affective tone, cognitive slowing, hopelessness, anhedonia, and social withdrawal before those changes are fully compressed into survey items. An LLM trained to read that language may therefore serve as a different kind of measurement instrument, one that is sensitive to latent severity rather than only to response options.
The clinical appeal is obvious for intensive longitudinal monitoring. Daily surveys can become burdensome, and burden can degrade compliance exactly when monitoring is most needed. A brief diary spoken or transcribed in everyday language could lower friction for patients while preserving repeated measurement across short intervals. For research studies that depend on dense symptom sampling, that could improve feasibility without abandoning clinical relevance.
Brief diaries may fit the demands of intensive longitudinal monitoring
These findings point toward a practical use case rather than a wholesale replacement for standard instruments. The performance of the model depended on diary transcripts and on a cohort already engaged in mental health assessment, so the context was structured and the signal likely benefited from a relatively high-information setting. Even so, the fact that LLM ratings aligned with both self-report and interview severity across days strengthens the case for language-based monitoring as a serious measurement strategy.
The result also raises an important conceptual issue for depression research. Self-report remains central because it is direct, standardized, and well understood, yet it is still constrained by item wording and respondent fatigue. Language models offer a different route: they can synthesize narrative content at scale and potentially detect shifts that survey scales do not explicitly ask about. In that sense, the model is not competing with questionnaires so much as reading a separate channel of symptom expression.
A one-sentence conclusion is hard to avoid here: daily language was informative enough to mirror clinically relevant depression patterns, and in some analyses it added information beyond the surveys themselves.
Still, the study leaves a number of operational questions open. The cohort was modest in size, and the article does not establish how well the approach would perform in broader community samples, across different speaking styles, or in settings with less frequent clinical contact. It also does not show whether the same model performance would hold when diaries are shorter, noisier, or less emotionally explicit. Before this approach moves toward routine use, future work will need to test durability across populations, settings, and collection formats, and to determine how such ratings should be interpreted when they diverge from self-report rather than reinforce it.